TL;DR

Argonne National Laboratory proof-of-concept that applies Microsoft SEAL’s CKKS FHE to a small fully-connected MNIST classifier, demonstrating feasibility while documenting roughly five orders of magnitude runtime overhead and large ciphertext blow-up [Abstract][§4]. The motivating application is privacy-preserving third-party analysis of proprietary reactor structural-health and nondestructive evaluation (NDE) data [§1].

Problem and motivation

Third-party / expert analysis is cost-effective for specialised reactor structural health monitoring and nondestructive evaluation work, but exposes proprietary data to privacy risk [§1, p. 4]. The authors propose using Fully Homomorphic Encryption so a client can hand encrypted data to a third-party processor that performs ML/AI inference without ever seeing the plaintext, then returns an encrypted result for the client to decrypt [Figure 1, p. 4]. The threat model is implicit: a client/server split where the server performs operations on ciphertext only; no explicit honest-but-curious / malicious distinction is stated [§1].

Key contributions

FHE setup

ML setup

Datasets

Dataset Task Size (train/test) Modality Notes
MNIST Handwritten 10-class digit classification 1000 train / 100 test (subset used for evaluation) 28x28 black-and-white images Used as proof-of-concept; “standardized” images [§4, p. 10]

Pipeline diagram

flowchart LR
    subgraph Client
        A[Load 28x28 image] --> B[Encrypt CKKS]
    end
    subgraph Server
        C[Plaintext weights] --> D[FC1 100 nodes]
        B -. ciphertext .-> D
        D --> E[Poly-ReLU deg 2]
        E --> F[FC2 10 nodes]
    end
    subgraph ClientDec [Client]
        G[Decrypt CKKS] --> H[Argmax label]
    end
    F -. ciphertext .-> G

Pipeline steps (text)

  1. Client loads a standardised 28x28 MNIST image [§4, p. 10].
  2. Client encrypts the input under CKKS using Microsoft SEAL at poly-modulus degree 8192 or 16384 [§2, p. 6].
  3. Client transmits the ciphertext to the third-party server [Figure 1, p. 4].
  4. Server applies the first fully-connected layer (100 nodes) using plaintext weights against the ciphertext [§3, p. 8].
  5. Server applies the degree-2 polynomial approximation of ReLU element-wise; rescales / relinearises after each multiplication [§2, §3, p. 6-8].
  6. Server applies the output fully-connected layer (10 nodes); SoftMax is not applied under encryption because argmax is preserved by the pre-image [§3, p. 8].
  7. Server returns the encrypted logit vector to the client [Figure 1, p. 4].
  8. Client decrypts and takes argmax to obtain the predicted digit [§3, p. 8].

Architecture diagram

flowchart LR
    I([Input 28x28=784]) --> FC1[FC 100]
    FC1 --> A1[poly-ReLU deg 2]
    A1 --> FC2[FC 10]
    FC2 --> O([argmax / SoftMax client-side])

CNN (plaintext only, for context)

flowchart LR
    I([Input 28x28]) --> C1[Conv]
    C1 --> C2[Conv]
    C2 --> P[MaxPool]
    P --> FC1[FC]
    FC1 --> FC2[FC]
    FC2 --> O([Output 10])

CNN layer widths and kernel sizes are not specified in the text [Figure 5, p. 8]. The CNN was not converted to an FHE implementation [§3, p. 8].

Results

Metric This paper Baseline Hardware
Plaintext FCNN accuracy (MNIST) 83.8% - Not reported [Figure 6, p. 10]
Approximated FCNN accuracy (poly-ReLU, plaintext eval) 76.1% (-7.7 pts vs FCNN) 83.8% (FCNN) Not reported [Figure 6, p. 10]
PCA + FCNN accuracy 85.9% 83.8% (FCNN) Not reported [Figure 6, p. 10]
CNN accuracy (plaintext only) 59.1% - Not reported [Figure 6, p. 10]
Plaintext FCNN inference time 2.001x10^-3 s per image - Not reported [Figure 7, p. 10]
HE (CKKS) FCNN inference time 295.3 s per image (~5 min) 2.001x10^-3 s plaintext Not reported [Figure 7, p. 10]
HE slowdown ~150,000x 1x plaintext Not reported [§4, p. 10]
Storage blow-up (10 images) 2.45 GB (N=8192) / 7.73 GB (N=16384) 68.4 KB plaintext Ratios 35,818x and 113,011x [Figure 2, p. 6]

The accuracy reported for the “Approximated Neural Network” in Figure 6 is described in §4 as identical between the plaintext-poly and HE-poly runs (“we receive an identical result between the NN and Approximated NN models” referring to outputs, not accuracy) [§4, p. 10] — the 76.1% number is therefore the accuracy of the network with poly-ReLU, whether evaluated in plaintext or under HE.

Limitations and assumptions

The paper cites but does not run head-to-head benchmarks against:

No quantitative comparison table against these systems is provided [§4-5].

Code and artifacts

Not released. No repository URL is given in the report [§1-5].

Extra diagrams (optional)

Threat model

flowchart LR
    C[Client] -- enc(image) --> S[Third-party Server]
    S -- enc(logits) --> C
    W[Plaintext weights] --> S

The server holds plaintext model weights; only the input/output are encrypted. No explicit adversary model is stated in the text [§1, p. 4-5].

Activation approximation

ReLU is approximated over [-20, +30] by p(x) = 2.593 + 0.4752x + 0.0173x^2 (degree 2), fit by minimising the integrated squared error [§2, Eq. 1-3, p. 7]. See Figure 3 in paper for the visual comparison [Figure 3, p. 7].

Open questions